[Final report of the visiting stay.] Assoc. Prof. Juraj Parajka. TU Vienna, Austria

Transcription

1 Validation of HSAF snow products in mountainous terrain in Austria and assimilating the snow products into a conceptual hydrological model at regional scale [Final report of the visiting stay.] Assoc. Prof. Juraj Parajka TU Vienna, Austria March 2014

2 Summary This report evaluates the mapping accuracy of MSG-SEVIRI snow products and its value for improving hydrological simulations. It has two sections. The first part evaluates the mapping accuracy of the MSG-SEVIRI operational snow cover product over Austria. The advantage of MSG-SEVIRI snow cover product is improved temporal resolution, but relatively lower spatial resolution of 5km. The snow mapping accuracy is examined at 178 stations with daily snow depth observations and compared with the daily MODIS combined (Terra + Aqua) snow cover product in the period April June The results of the accuracy assessment show that the 15-minute temporal sampling allows a noticeable reduction of clouds in the snow cover product. The mean annual cloud coverage is less than 30% in Austria, as compared to 52% for the combined MODIS product. The mapping accuracy for cloud-free days is 89% as compared to 94% for MODIS. The largest mapping errors are found in regions with large topographical variability. The errors are noticeably larger at stations with elevations that differ much from those of the mean MSG- SEVIRI pixel elevations. The median of mapping accuracy for stations with absolute elevation difference less than 50m and more than 500m is 98.9% and 78.2%, respectively. A comparison between the MSG-SEVIRI and MODIS products indicates an 83% overall agreement. The largest disagreements are found in alpine valleys and flatland areas in the spring and winter months, respectively. The second part of the report investigates the integration of MSG-SEVIRI snow cover in the calibration of a conceptual hydrologic model. The accuracy of traditional (to runoff only) and multiple objective calibration is evaluated over 144 catchments of different size and catchment characteristics in Austria. The results indicate that assimilation of MSG-SEVIRI snow cover data in model calibration allows a robust estimation of hydrologic model parameters. The runoff model accuracy of multiple of multiple objective calibration is similar or only slightly poorer then obtained by calibration to runoff only (traditional single-objective model calibration). However, the snow model efficiency is slightly improved, particularly in catchments with lower elevations (mean elevation below 900 m a.s.l.) and in catchments with absent or sparse precipitation measurements. Interestingly, the runoff and snow model accuracy is not improved in the high alpine catchments. This findings is likely related to larger MSG-SEVIRI snow cover mapping errors in the mountains, which is caused by significant topographical variability and coarser spatial resolution of the MSG-SEVIRI.

3 1 Introduction and objectives Snow is an important component of the water cycle and of the climate evolution. Monitoring snow parameters (e.g. snow-covered area and snow-water equivalent) is challenging work for meteorologists and climatologists who are studying climatic and atmospheric variability globally. Due to its high albedo, snow also plays an important role in the Earth s energy balance, affecting weather and climate; thus snow influences the evolution of weather and climate. The accurate observation of the spatial and temporal variability of snow-covered area is important for monitoring global climate change. In hydrology, snow acts as a highvolume water storage controlling water reservoirs, affecting the evaporation process. Snow melting constitutes a potential risk of flooding in certain areas, but it is also an excellent source of energy for power plants. Therefore monitoring and estimating the snow parameters play an important role in predicting discharges during melting seasons. Snow height observations are available over large areas on many meteorological stations. These data are very dependent on the local conditions (wind, vegetation, slope, aspect, etc). Especially for mountainous areas the scarcity of the field observations and the representativeness of the stations for the areal extent due to the complexity of the terrain make the use of ground observations in snow monitoring difficult and insufficient. For mountainous regions, satellite imagery is the most convenient way for keeping track of snow cover extent considering the inaccessibility due to the difficulties of rough terrain and high elevations. Remote sensing data has been used for better comprehension of information on snow cover extent (Painter et al., 2003, Cline et al., 1997). Several satellite sensors have been used for snow cover mapping such as: AVHRR, MODIS, and MERIS (Harrison and Lucas 1989, Hall et al., 2002, Tampellini et al., 2004). MODIS has a good temporal and spatial resolutions for snow cover monitoring, therefore it has been utilized in numerous studies (Parajka and Blöschl, 2012). Various studies have been done on the validation of MODIS snow cover products under a variety of snow and land cover conditions. Most studies show an overall clear sky accuracy of 94% compared with ground measurements (Hall and Riggs, 2007; Parajka and Blöschl, 2006; Parajka and Blöschl, 2012). Lower accuracies are typically obtained in the fall and spring and under thin-snow conditions and in dense forested areas.

4 Frequent cloud coverage is another problem in snow mapping through remote sensing. Cloud cover discrimination is the most challenging problem in retrieving snow cover information from the satellite images acquired in the optical portion of the spectrum. Better temporal resolution satellite data improve the discrimination of cloud. The possibility of 37% cloud cover reduction by merging 15min observations from MSG-SEVIRI (H10 product) compared to using only one daily observation from MODIS has been obtained (Sürer and Akyurek, 2012). The satellite snow observations are being used in the field of hydrology (Andreadis and Lettenmaier, 2005; Rodell and Houser, 2004; Su et al, 2008; Zaitchik and Rodell, 2008; Bavera and De Michele, 2009; Tekeli et al., 2007). Remotely sensed snow covered area information has been used successfully in snowmelt and runoff models (e.g. Yang et al., 2003, Clark et al., 2006, Dressler et al., 2006, Kolberg and Gottschalk, 2006, Kolberg et al., 2006, Andreadis and Lettenmaier, 2006, Udnaes et al., 2007, Parajka and Blöschl, 2008). Remotely sensed snow water equivalent has also been used in some studies (e.g. Derksen et al. 2003, Andreadis and Lettenmaier 2006, Pulliainen, 2006). Remotely sensed snow cover information can be either used as direct input into a hydrological model (Tekeli et.al., 2005) or the simulated snow water equivalent values can be compared with the snow cover data indirectly (Parajka and Blöschl, 2008). Ideally, a system that optimally combines snow information from both, remote sensing and modeling predictions and at the same time accounts for the limitations of each, should provide estimates that are superior to those derived from either models or remote sensing alone. This method is commonly known as data assimilation (McLaughlin, 1995). There are two types of snow assimilation in hydrologic models. The first refers to the constraining of hydrologic models in model calibration procedure (i.e. Parajka et al., 2006, Parajka and Blöschl, 2008). The latter refers to a diagnostic snow assimilation to directly constrain the snow state variable of the hydrologic model (typically snow water equivalent). There are different assimilation methods that have been applied to land surface models in order to update the snow information. Recently, some of those methods have been utilized in hydrological modelling studies as well. Among these methods, variants of Kalman Filter technique are the most preferred ones. It is the method of adjusting uncertain variables and parameters in order to obtain the best fit to the values from observations (Houser et. al.,

5 1998). Direct insertion method is another option to update and assimilate snow information in hydrological models as Liston et. al. (1999) has successfully applied in a regional climate model for snow association and Rodell et. al. (2004) used this method to assimilate MODIS data into a global land surface model. Statistical interpolation technique is sort of an improvement for direct insertion which is applied by Brasnett (1999) to assimilate snow depth observations from synoptic stations. Thirel et al. (2011) compared Ensemble Kalman filter and particle filter assimilation techniques to improve the runoff simulation with a spatially distributed hydrological model. Their results and discussion to that paper indicate that there is still more research needed for better understanding on how to robustly assimilate satellite snow cover data into hydrologic models. The robust diagnostic assimilation however requires a three step assessment. The first step includes the analyses of the accuracy and uncertainty of satellite snow cover product. The second step evaluates the accuracy and sensitivity of hydrologic model to the selected snow characteristic (i.e. snow cover, or snow water storage). The third step is then a final assessment of different diagnostic assimilation procedures. The main objectives of the visiting stay and the subsequent report are is to assess and summarize the results of the two first steps: 1) Validation of snow cover (H10) and snow water equivalent (H13) satellite products over mountainous part of Austria (Section 2), 2) Assimilation of H10 product to a hydrological model calibration and evaluating the runoff model performance over large sample of catchments in Austria (Section 3). The results of this evaluation can serve as a basis for selecting and building a robust diagnostic assimilation in the next future.

6 2 Validation of snow cover (H10) and snow water equivalent (H13) satellite products over mountainous part of Austria Monitoring and modeling of snow characteristics is important for many hydrological applications, including snowmelt runoff forecasting and water resources assessment using a range of techniques (e.g. Blöschl and Kirnbauer, 1991; Blöschl et al., 1991; Nester et al., 2012). The large spatial variability of snow cover, particularly in mountains, limits the use of ground based snow observations. Satellite imagery is thus an attractive alternative, as the resolution and availability does not depend much on the terrain characteristics (Parajka and Blöschl, 2008). Recently, operational satellite products have become available that provide snow cover information at different spatial and temporal resolutions (Table 1). Table 1 indicates that most of the current products provide daily snow cover information at spatial resolutions ranging from 500 m to 5 km. The snow mapping accuracy with respect to ground snow observations for cloud free conditions varies between 69% and 94% in the winter seasons. The main limitation of existing optical platforms operating at daily time scale is cloud coverage, which significantly reduces the availability of snow cover information. There exist different approaches for cloud reduction including space-time filtering (e.g. Parajka and Blöschl, 2008, Gafurov and Bárdossy, 2009, Hall et al., 2010, among others), but clouds are real and the accuracy of such approaches decreases with their efficiency to reduce clouds. An alternative to the space-time filtering of daily products is to merge satellite images obtained at higher temporal resolution. The new generation MSG-SEVIRI product provides snow cover information at 15 minutes temporal resolution for the whole northern hemisphere. The preliminary assessment of data from one snow season over Eastern Turkey (Surer and Akyurek, 2012) indicates that merging of 32 consecutive images per day enables a 37% reduction of clouds in comparison to the MODIS daily product and improves the mapping of regional snow-cover extent over mountainous areas.

7 Table 1. Summary of some existing operational satellite snow cover products. Snow cover product Sensor Available Spatial Tempora Mapping accuracy since resolutio l n resolutio n NOHRSC/ NOAA/AVH RR+GOES km Daily 76% (Klein and Barnett, 2003) NOAA/NESDIS (IMS) GOES+SSM/ km Daily, 85% (Romanov et I weekly al., 2000); <20%(October), ~60% (November), ~95%(December), ~70%(March) (Brubaker et al., 2005) MOD10A1,MYD10A1, MOD10A2,MYD10A2, MOD10C1,MYD10C1, MOD10CM,MYD10CM MODIS- Terra/Aqua 2000/ m o Daily, 8-day, monthly ~94% (see e.g. Hall and Riggs, 2007 or summary in Parajka and Blöschl, 2012) HSAF (Eumetsat) MSG-SEVIRI km Daily 80% compared to IMS (Siljamo, and Hyvärinen, 2011); 69-81% in winter months (Surer and Akyurek, 2012)

8 The main objective of this part of the report is to assess the accuracy of the new MSG-SEVIRI snow products over Austria in the period The spatial and temporal variability in mapping accuracy is examined for a large number of meteorological stations observing snow depth and evaluated against combined MODIS snow cover product. Austria is an ideal test bed for such an assessment, as it allows evaluating the mapping accuracy in different altitudinal zones ranging from the lowlands to the high Alpine environment. 2.1 MSG-SEVIRI snow cover product (H10) The Spinning Enhanced Visible and Infrared Imager (SEVIRI) is an optical imaging radiometer mounted on board of the geostationary Meteosat Second Generation (MSG) satellite operated by EUMETSAT. MSG-SEVIRI provides continuous imaging of the Earth in 12 spectral channels with a repeat cycle of 15 min. The imaging spatial resolution is 3 km at sub-satellite point (Aminou, 2002) and degrades to 5km over Europe. The snow cover mapping is based on a multi-channel retrieval algorithm. It exploits the high reflectivity of snow in the visible spectrum and the low reflectivity at shorter wavelengths. The snow cover retrieval algorithm differs for flat and mountain regions. In flat regions, the algorithm utilizes the top-of-atmosphere radiance of 6 SEVIRI channels (0.6, 0.8, 1.6, 3.9, 10.8 and 12.0 µm) and brightness temperatures of three channels (3.9, 10.8, and 12.0 µm). The snow recognition is based on the snow cover classification (Siljamo and Hyvarinen, 2011). The cloud-snow discrimination for flatlands relies on the cloud mask (CMa) provided by the Nowcasting and Very Short Range Forecasting Project (NWCSAF, 2007). In this product clouds are classified only into two classes (cloud contaminated and cloud filled). In the mountains, the snow recognition algorithm uses the snow index (SI) which relates 0.6µm, and 1.6µm SEVIRI channels. The cloud-snow discrimination is based on the CMa and cloud type (CT) product of the NWCSAF. The CT product has 15 different cloud types which allow more robust cloud recognition (Surer, 2008). Both algorithms use sun zenith angle for discarding the low illuminated areas, and land surface temperature values for covering all cold pixels below freezing point (Romanov et al., 2003). The main difference in the algorithms is the location of the samples collected for developing the thresholding method and the cloud-snow discrimination applied in the retrieval. A detailed description of the

9 MSG-SEVIRI snow algorithm is presented in the Algorithm Theoretical Basis Document (HSAF, 2010). The definition of the mountainous areas is based on the mean altitude and standard deviation of the slope within 5 km x 5 km pixels (Lahtinen et al., 2009). The area is defined to be mountainous if the mean altitude exceeds 1000 m or the mean altitude exceeds 700 m, and the standard deviation of the slope is greater than 2 o or range in mean altitude exceeded 800 m and mean altitude exceeds 500 m. Daily snow cover maps are derived from 32 images per day, blending data from 08:00-15:45. Snow cover is mapped when there are at least 4 hits of snow recognition in a day. The final snow cover product, which is merged at Finnish Meteorological Institute, has snow, land, cloud, water and unclassified classes. An example map for Europe is presented in Figure 1. Figure 1. Example of a MSG-SEVIRI snow cover map for February, 21st, Snow water equivalent product (H13) The snow water equivalent product (H13) is fundamentally based on the AMSR-E microwave radiometer measurements being flown on EOS-Aqua. The final estimates of snow water equivalent (SWE) represent the result of an assimilation process. The basic (very sparse) ground network of snow depth observation provides a first guess field that is converted into MW brightness temperatures by an snow emission model developed by Pulliainen et al., (1999). The assimilation process forces the first guess field to optimally match the AMSR-E

10 brightness temperatures (PUM-13, 2012). The snow emission model describes the spaceborne observed microwave brightness temperature as a function of snow pack characteristics and by considering the effects of atmosphere, forest canopy and land cover category (fractions of open and forested areas). The retrieval algorithm slightly differs for flat or forested area and for mountainous regions. In the mountains, a different relationship for estimating pixel snow density and an interpolation, which also accounts for ancillary data (i.e. DEM, aspect and slope maps) is applied. A detailed description of the model and its performance is given in PUM-13 (2012). The final product represents gridded estimates of SWE in 25km spatial and daily temporal resolution. An example for Austria is presented in Figure 2. Figure 2. An example of gridded snow water equivalent as estimated by the H13 product in Austria. 2.3 Study area and snow cover data The accuracy of snow products is evaluated over Austria. Austria is located in the temperate climate zone, where the Alps act as a dominant barrier between continental climate in the north and the meridional circulation from the Adriatic Sea in the south. Elevations range from 115m in the flatlands to more than 3700 m in the mountains (Figure 3). Mean annual precipitation varies between 400 mm in the eastern flatlands and almost 3000 mm in the

11 western part of the Alps. The mountainous parts of Austria are covered by snow for several months in a year (Parajka and Blöschl, 2006), while the flatlands are characterized by warm and dry summers and cold winters without significant snowfall. Land use is mainly agricultural in the lowlands and forest in the medium elevation ranges. Alpine vegetation and rocks prevail in the highest catchments. Figure 3. Topography of Austria and location of 178 stations with daily snow depth measurements in the period April June Red and blue colors represent meteorological stations located in the flatland and (81 stations) and mountain (97 stations) regions according to the MSG-SEVIRI mountain mask, respectively. Table 2. Number of meteorological stations in different elevation zones. Elevation zone (m a.s.l.) Number of stations

12 Snow cover data used for MSG-SEVIRI evaluation includes ground snow depth measurements at 178 meteorological stations (Figure 3) and daily MODIS satellite snow cover images from April June The snow depth readings are taken from permanent staff gauges and represent point measurements performed daily at 07:00 AM with 1 cm reading precision (Parajka and Blöschl, 2006). Table 2 summarizes the number of stations in different elevation zones and indicates that most of the stations are located in elevation zones between 500 m and 1000 m. In the mountains, the stations tend to be located at lower elevations, typically in the valleys, which suggest a slight bias of the validation statistics towards lower elevations. The satellite snow cover images have been acquired by the MODIS instrument mounted on Terra and Aqua satellites of the NASA Earth Observation System. The daily Terra (MOD10A1, V005) and Aqua (MYD10A1, V005) snow products are available through the Distributed Active Archive Center located at the National Snow and Ice Data Center (NSIDC, The spatial resolution of the products is 500m. For the validation, the snow cover product obtained from the Terra satellite and a combined product of the Terra and Aqua satellites are used. The two products are combined to reduce cloud coverage in the mountains (Parajka and Blöschl, 2008). In the combined product, the pixels classified as clouds in the Terra images are updated by the Aqua pixel value of the same location if the Aqua pixel is snow or land. This approach combines satellite observations on the same day, shifted by several hours. 2.4 Methodology of snow product accuracy assessment Evaluation of the snow mapping accuracy is performed in two steps. In the first step, the accuracy of H10 and H13 products is evaluated at meteorological stations by using daily snow depth observations. Snow depth observations at the stations are considered as ground truth for each H10 and H13 pixel that is closest to each station. The ground is considered as snow covered if the snow depth measurement exceeds 1 cm. In the second step, H10 images are compared with daily MODIS snow cover maps. In this case, the frequency of MODIS snow, no snow and cloud classes is estimated and compared within each H10 pixel.

13 The snow cover mapping accuracy with respect to snow depth observations is quantified by three variants of the accuracy index: k A, k M and k C. The overall accuracy index k A is estimated at each meteorological station and compares the sum of all correctly classified days with the presence of snow and no snow to the number of all cloud-free days at each meteorological station (station-days) in the selected period. The seasonal accuracy index k M is defined in a similar way, but relates the sum of all correctly classified station-days (snow-snow, no snowno snow) at different meteorological stations to the number of all cloud-free station-days at those stations in a particular month. The k M index is estimated separately for all stations located in the mountain and flatland areas as defined by the MSG-SEVIRI (H10) mountain mask (Figure 3), respectively. The all-days accuracy index k C relates the correctly classified station-days to the total number of station-days in the selected period, including days with cloud cover. It is also estimated for each month and two groups of stations (mountain and flatland). Additionally to the three accuracy indices, two types of mapping errors are quantified with respect to the ground snow depth observations: the H10 and H13 misclassification of land as snow termed here the snow overestimation error (k O ) and the misclassification of snow as land termed the snow underestimation error (k U ). Both types of errors relate the sum of misclassified station-days to the total number of station-days in each particular month and mask region. The agreement between H10 and MODIS snow cover products is quantified by the index of overall m A and seasonal agreement m M. These indices are defined in a similar way as the k A and k M, but instead of using snow depth observations at meteorological stations, the aggregated frequencies of MODIS snow, land and cloud classes within each MSG-SEVIRI (H10) pixel are used. The comparison is performed at the coarser spatial resolution of the H10 and for those H10 pixel-days where the relative frequency of MODIS pixels classified as clouds is less than 60%. Our test simulations (not shown here) indicate that the results are insensitive to the selection of this threshold between 40 and 70%. In the m A and m M evaluation, the ground is considered as snow covered if the frequency of MODIS snow pixels within the H10 pixel is at least 50% of the sum of MODIS pixels classified as snow and land. The presence of no snow (land class) is considered in the same way, i.e. the frequency of MODIS pixels classified as land is larger than the sum of snow and land pixels.

14 2.5 Validation of H10 snow cover product against ground snow depth measurements The snow cover accuracy (k A ) of MSG-SEVIRI (H10) estimated for cloud-free days at the meteorological stations is presented in Figure 4 and summarized in Table 3. The k A varies between 51.3% at the Villacher Alpe (2140 m a.s.l.) in the Eastern Alps (Carinthia) and almost 100% in Gross-Enzersdorf (154 m a.s.l.) near Vienna. Table 3 indicates that the MSG-SEVIRI accuracy is larger in the flatland than in the mountain regions, i.e. the median of k A is 98.8% and 84.3% in the flatland and mountain regions, respectively. Table 3. Overall accuracy k A (%) of the MSG-SEVIRI (H10) snow cover product for cloud-free days at the meteorological stations. Stations in flatland and mountains are stratified according to the mountain mask used for the H10 product (Figure 3). Stations in Statistics All stations mountains Stations in flatland Count Minimum k A % percentile k A % percentile k A % percentile k A Maximum k A

15 Figure 4. MSG-SEVIRI (H10) snow product overall accuracy k A (%) at 178 meteorological stations in the period April 2008-June Figure 5 shows a clear decrease of snow mapping accuracy with increasing elevation of the meteorological stations. The results indicate that this tendency is caused mainly by increasing sub-grid topographical variability in the mountains. Meteorological stations are often situated at different elevations than the mean elevation of MSG-SEVIRI pixels, which causes biases between station and satellite snow cover observations. As is indicated in the left panel of Figure 5, the mapping accuracy is larger for stations with smaller elevation difference. For example, the median of k A for stations with absolute elevation difference less than 50m and more than 500m is 98.9% and 78.2%, respectively. For the station with the largest mapping errors (Villacher Alpe) the elevation difference is larger than 960m. The stations located significantly below or above the pixel mean may have noticeably different snow cover observations (right panel of Figure 5). The snow cover observations at meteorological stations in Austria show a clear linear relationship (R 2 =88%) between snow cover duration and the altitude, indicating an increase of snow cover duration by 2.8%/100m (not shown here). An elevation difference of 500 m can therefore be easily transferred in about 14% difference in snow cover duration and thus different snow cover mapping accuracy. Interestingly, the MSG-SEVIRI mapping accuracy is larger than 90% for two stations situated above 2000 m a.s.l. (Ischgl-Idalpe and Pitztaler Gletscher), but located

16 approximately at the mean elevation of the MSG-SEVIRI pixel. This finding indicates the importance of the spatial resolution and sub-grid topographical variability for the assimilation of satellite snow cover images in operational hydrological applications. Figure 5. Relationship between MSG-SEVIRI (H10) snow mapping accuracy (k A ) and elevation of the meteorological stations. Color of the triangles in the left panel indicates the difference between elevation of the meteorological stations and mean elevation of the respective MSG-SEVIRI pixels (as derived from a 25m digital elevation model). Color of the symbols in the right panel shows relative snow cover duration observed at the meteorological stations in the period April 2008-June The seasonal frequencies of MSG-SEVIRI snow mapping accuracy (k M ) is presented in Figure 6. The results show that, in the mountains, the k M accuracy varies between 70-77% in the winter and 92-97% in the summer months. The flatland region has typically much shorter snow coverage, which most likely results in larger k M accuracy between April and October, but larger mapping errors (k M between 79 and 83%) in the winter months. As compared to

17 k M obtained for the MODIS/Terra and MODIS/combined snow cover products, the MSG- SEVIRI mapping accuracy is 10-13% lower in the mountains and 3-11% in the flatland area in the winter months. However, the MSG-SEVIRI product contains significantly less pixels classified as clouds than MODIS, particularly in the mountains (Figure 7, top panels). ). Here, the merging of 32 MSG-SEVIRI images per day reduces cloud coverage between 15 to 29% in the period November-June as compared to MODIS-combined product. The cloud reduction is even about 7% larger when compared to the MODIS-Terra product. In the period July- October, the cloud coverage of MSG-SEVIRI is similar to that of MODIS in the mountains. Interestingly, in the flatland areas a decrease in cloud coverage is observed only in the period April and September. In the winter months, MSG-SEVIRI indicates cloud coverage larger than 75%, which is similar or even slightly larger than indicated by the MODIS products. This is probably caused by use of different cloud masking algorithms. Figure 6. Seasonal frequency of snow mapping accuracy k M for the MSG-SEVIRI (H10), MODIS-Terra and MODIS-combined products estimated for cloud-free days in the period April June Left and right panels show the results for meteorological stations in the mountain (97 stations) and flatland (81 stations) regions, respectively. The reduction in clouds, particularly in the mountains, then translates into an improvement of all-days mapping accuracy k C (Figure 7, bottom panels). The k C accuracy assumes clouds as

18 a mapping error and it varies for MSG-SEVIRI between 26-31% (mountains) and 9-25% (flatland areas) in the winter and spring periods. In the mountains, this is about 3-14% larger than the k C obtained for the MODIS dataset. In the flatland areas, the large cloud coverage in winter does not enable an increase in k C as compared to MODIS products. The evaluation of k C clearly indicates the tradeoff between increased cloud reduction due to higher temporal sampling (32 images per day) and higher mapping error due to coarser spatial resolution (particularly in the mountains) of the MSG-SEVIRI snow product. Figure 7. Seasonal frequency of the clouds (top panels) and snow mapping accuracy k C for the MSG-SEVIRI (H10), MODIS-Terra and MODIS-combined products estimated for all days in the period April June Left and right panels show the results for meteorological stations in the mountain (97 stations) and flatland (81 stations) regions, respectively.

19 The seasonal frequency of MSG-SEVIRI mapping errors is summarized in Table 4. Table 4 compares the overestimation (k O ) and underestimation (k U ) errors of MSG-SEVIRI, MODIS- Terra and MODIS-combined datasets as observed at meteorological stations. The general distribution of MSG-SEVIRI errors shows a typical seasonal pattern of larger errors in winter and spring and smaller errors in summer. In comparison to MODIS products, the MSG-SEVIRI mapping errors are particularly larger during the snowmelt season in the mountains (4-9%) and somewhat larger during the winter months in the flatlands (1-3%). Table 4. Seasonal frequency of overestimation (k O ) and underestimation (k U ) mapping errors (%) estimated for the MSG-SEVIRI (H10), MODIS-Terra and MODIS-combined snow cover products in the period April June The mapping errors are estimated at 97 and 81 meteorological stations in the mountain (Mnt) and flatland (Flat) areas, respectively. (The largest mapping error for each product and mask area is marked by bold print). Season MSG-SEVIRI MSG-SEVIRI MODIS-Terra MODIS-Terra MODIS-comb. MODIS-comb. overest. k O underst. k U overest. k O underest. k U overest. k O underest. k U (Mnt/Flat) (Mnt/Flat) (Mnt/Flat) (Mnt/Flat) (Mnt/Flat) (Mnt/Flat) January 4.6/ / / / / /1.2 February 4.3/ / / / / /0.8 March 6.1/ / / / / /0.6 April 8.8/ / / / / /0.2 May 5.5/ / / / / /0.0 June 2.2/ / / / / /0.0 July 1.3/ / / / / /0.0 August 0.9/ / / / / /0.0 September 1.0/ / / / / /0.0 October 4.0/ / / / / /0.0 November 6.1/ / / / / /0.3 December 5.1/ / / / / /0.6

20 A detailed analysis of k O and k U errors (Figure 8) indicates that the MSG-SEVIRI mapping errors are much larger at stations that are located at different elevations than the mean elevation of the closest MSG-SEVIRI pixel. The largest k O, i.e. more than 25% in April or 15% in November, is estimated at stations that are located more than 500 m lower than the pixel mean. Similarly, the largest k U errors are found at stations located more than 500m above the pixel mean. The evaluation of MSG-SEVIRI mapping errors at stations that are located at approximately the same elevation (yellow triangles in Figure 8) indicates that the MSG- SEVIRI tends to more frequently underestimate snow cover in winter than overestimating it. The largest k O errors are less than 0.5%, but k U errors exceed 3% in the winter months. Figure 8. Seasonal frequency of MSG-SEVIRI (H10) snow overestimation (k O, left panel) and underestimation (k U, right panel) errors summarized for stations with different elevation difference between meteorological station and respective MSG-SEVIRI pixel mean. The elevation difference is estimated as station elevation minus mean pixel elevation (as derived from a 25m digital elevation model).

21 2.6 Comparison between H10 and MODIS snow cover products The overall agreement between the MSG-SEVIRI and MODIS-combined maps (m A ) is summarized in Table 5. The m A vary between 57.3% to 92.7%, with a median of 82.5%. The difference in medians between the flatland (82.9%) and mountain (81.6%) regions is not large. The spatial patterns indicate (Figure 9) that m A is between 80% and 90% in the flatland, with an exception in the hilly region at the border between Upper and Lower Austria (Waldviertel), where m A is less than 75%. In the mountains, the m A variability tends to be larger. The m A agreement is over 90% in the high mountain locations, but smaller than 65% in the Alpine valleys in western Austria. It is also less than 70% in the south-eastern part of the (Lahtinen et al., 2009) mountain mask region (Styria). The relationship between m A and altitude is plotted in Figure 10. While in the flatlands, m A tends to decrease with elevation, in the mountains there is a tendency of increasing m A with altitude. The results show that the largest m A variability in Austria is in the regions with altitudes between 700 and 1500 m. Table 5. Overall agreement m A (%) between MSG-SEVIRI and MODIS-combined snow cover products for MSG-SEVIRI pixels with less than 60% MODIS cloud coverage. The agreement m A accuracy is evaluated for all MSG-SEVIRI pixels, flatland and mountain mask areas in Austria. Statistics All pixels Pixels in mountains Pixels in flatland Count Minimum m A % percentilemk A % percentile m A % percentile m A Maximum m A

22 Figure 9. Overall accuracy k A of MSG-SEVIRI (H10) with respect to the MODIS- combined product in the period April June k A is estimated for the MSG-SEVIRI pixels where MODIS cloud coverage is less than 60%. Pixels with black outline indicate the MSG-SEVIRI mountain mask. Figure 10. Relationship between mean MSG-SEVIRI (H10) pixel elevation and the overall agreement (m A ) between the MSG-SEVIRI (H10) and MODIS-combined products. Red and blue points represent MSG-SEVIRI pixels in the flatland and mountain mask areas, respectively.

23 The seasonal variability (m M ) in the agreement between MSG-SEVIRI and MODIS is presented in Figure 11. In the flatland areas (red line), m M is the largest in April and July and less than 70% in the winter months. The m M amplitude is smaller in the mountains (blue line), ranging from more than 85% in May, June and August to 70% in September. Figure 11. Seasonal agreement m M between the MSG-SEVIRI (H10) and MODIS- combined products for MODIS cloud-free pixels in the period April June Red and blue lines represent mountain and flatland areas, respectively. A more detailed evaluation of the spatio-temporal patterns of the agreement between MSG- SEVIRI and MODIS is presented in Figures 12 and 13. Figure 12 compares the spatial patterns of the frequencies of three MSG-SEVIRI and MODIS mapping classes - clouds, snow, no snow. It is clear, that the agreement between the snow cover products is the largest for mapping the clouds, for mapping the land in the flatland and snow in the high alpine areas. These cases occur in more than 25% of days in the selected period, in most of the MSG-SEVIRI pixels. The MSG-SEVIRI maps snow, while the MODIS combined product indicates clouds in 10-15% of days in the Alps. Interestingly, in the flatland, there are only a few days when both MSG-SEVIRI and MODIS indicate snow. The spatial patterns of the disagreement between the products, i.e. MSG-SEVIRI maps no presence of snow (land), but MODIS indicates snow, show that most of the cases are in Upper Austria, Styria and the mountain valleys. An opposite case occurs quite frequently in the mountain valleys of western Austria, where

24 MSG-SEVIRI and MODIS map snow and land in 10-15% of days, respectively. Figure 13 shows that MSG-SEVIRI overestimates snow in comparison to MODIS (middle panels) mainly in the summer for both mountain and flatland areas. The bottom panel (Figure 13) indicates that the opposite case, i.e. MSG-SEVIRI underestimates snow in the winter, is less frequent (up to 10%). There is quite a large frequency of days where MSG-SEVIRI maps land and MODIS indicates clouds. These cases occur in more than 20% of the days of each month in the flatland area. In the mountains, the reduction of clouds is noticeable in the winter months, where MODIS indicates clouds, but MSG-SEVIRI maps snow in more than 15% of the days. Figure 12. Relative frequency of days with agreement and disagreement between the MSG- SEVIRI (H10) and MODIS-combined snow cover products in the period April June 2012.

25 Figure 13. Mean seasonal frequency of days with agreement and disagreement between the MSG-SEVIRI (H10) and MODIS-combined snow cover products in each month of the period April June 2012.

26 2.7 Validation of H13 snow water equivalent product against ground snow depth measurements The validation of snow water equivalent (H13) product against ground snow depth measurements has two parts. The first part shows the ability of H13 to indicate if the region(pixel) is covered by snow. In the second part, the H13 estimates of snow water equivalent (SWE) is compared with SWE derived from snow depth observations at meteorological stations. The assessment of the overall snow cover accuracy (k A ) of H13 is presented in Figure 14. The k A varies between 24.4% at the Sonnblick (3109 m a.s.l.) in the Eastern Alps (Carinthia) and 99.6% in Seibersdorf (185 m a.s.l.) near Vienna. Similarly as for H10, the H13 accuracy is larger in the flatland than in the mountain regions, i.e. the median of k A is 88.8% and 80.3% in the flatland and mountain regions, respectively. The results indicate that in the Alps, the snow cover accuracy of H13 tends to be lower than H10 while in the flatland region (with shorter snow cover occurrence) is the H13 accuracy similar to H10 (presented in section 2.5). Figure 14. Overall accuracy (k A, %) of snow water equivalent (H13) product at 178 meteorological stations in the period April 2008-June 2012.

27 Figure 15. Relationship between snow mapping accuracy (k A ) of the snow water equivalent product (H13) and elevation of the meteorological stations. Color of the triangles in the left panel indicates the difference between elevation of the meteorological stations and mean elevation of the respective H13 pixels (as derived from a 25m digital elevation model). Color of the symbols in the right panel shows relative snow cover duration observed at the meteorological stations in the period April 2008-June The comparison of H13 snow cover accuracy with respect to the altitude of meteorological stations and the elevation difference between the stations and H13 pixel means is presented in Figure 15. The results show similar patterns as the accuracy assessment of H10 product. There is a clear decrease of snow mapping accuracy with increasing elevation of the meteorological stations, which confirms the effects of sub-grid topographical variability on the mapping accuracy. As is indicated in the left panel of Figure 15, the H13 snow cover estimates are closer to snow observations at stations, which are situated at the mean pixel elevation (i.e. snow cover mapping accuracy of H13 is larger for stations with smaller elevation difference between station and pixel mean). For example, the median of k A for stations with absolute elevation difference less than 50m and more than 500m is 91.3% and

28 81.3%, respectively. The k A difference between these groups of stations is, however, for H13 smaller than for H10. The H13 accuracy is even significantly larger for stations which are located significantly below pixel mean (i.e. more than 500 m), in comparison to H10. This is likely related to the underestimation of snow by the H13 snow product, which hence result in better representation of snow coverage at lower stations. The seasonal frequencies of H13 snow mapping accuracy (k M ) and its comparison to H10 and MODIS combined product is presented in Figure 16. The results show that, in the mountains, the k M accuracy varies between 38-56% in the winter and 97-98% in the summer months. The flatland region has typically much shorter snow coverage, which most likely results in almost perfect agreement between May and October, but larger mapping errors (k M between 51 and 56%) in the winter months. As compared to k M obtained for the MSG-SEVIRI (H10) in the winter months, the H13 mapping accuracy is 6-9% lower in the flatland, but significantly lower (21-38% ) in the mountains. In the summer months, the k M of H13 product is similar (flatland region) or even somewhat larger than the k M of H10. It should be noted, however, that Figure 16 compares the k M of H13 product with the cloud-free cases of H10 product, so the frequency of available information of H13 is larger than H10. A comparison of the all-days accuracy k C (Figure 17) clearly show, that the clouds in MSG- SEVIRI and MODIS product significantly reduce the amount of information for snow cover mapping. This also indicate a clear potential of merging microwave H13 and optical H10 products for further reduction of cloud frequency in MSG-SEVIRI.

29 Figure 16. Seasonal frequency of snow mapping accuracy k M for the H13, MSG-SEVIRI (H10) and MODIS-combined products estimated for cloud-free days in the period April June Left and right panels show the results for meteorological stations in the mountain (97 stations) and flatland (81 stations) regions, respectively. Figure 17. Snow mapping accuracy k C for the microwave H13, MSG-SEVIRI (H10) and MODIScombined products estimated for all days in the period April June Left and right panels show the results for meteorological stations in the mountain (97 stations) and flatland (81 stations) regions, respectively.

30 The seasonal frequency of H13 snow cover mapping errors is presented in Figure 18. The left and right panels (Figure 18) show the frequency of H13 over- (k O ) and under- (k U ) estimation errors, respectively. From this assessment, it is clear that the microwave H13 product significantly underestimate the snow cover at meteorological stations, particularly at locations, which are situated above the H13 pixel mean. The mapping errors are obviously the largest in winter and exceed 50% even for stations, which are located below the mean pixel elevation. The largest k U errors exceed 80% at stations situated more than 500m above mean pixel elevation in April and December. The snow cover over-estimation errors are very small, the largest k O errors exceed 20% only at locations significantly (more than 500m) below the mean pixel elevation in February and March. Figure 18. Seasonal frequency of H13 snow overestimation (k O, left panel) and underestimation (k U, right panel) errors summarized for stations with different elevation difference between meteorological station and respective H13 pixel mean. The elevation difference is estimated as station elevation minus mean pixel elevation (as derived from a 25m digital elevation model).

31 A clear underestimation of snow is documented also in Figures 19 and 20, which compares H13 microwave estimates of SWE with SWE derived from daily snow depth observations at two meteorological stations in the mountains (Figure 19) and flatland (Figure 20) regions. The station SWE is plotted as a range of SWE values by using two different snow densities (0.150 and 0.300kg/m 3 ) in the derivation of SWE from the snow depth values (SWE=product of snow depth and snow density). Even if the snow density is not measured on daily time scale, the range of values used in the assessment shows a probable variability during the snow seasons. The results indicate that the maximum SWE values from the H13 product are around 100mm, which is significantly lower than derived from the observed snow depth measurements. The snow depth measurements at Brand station (Figure 19) exceed in some days 5m, which is clearly not captured by the H13 product. Also the snow depth observations at the beginning of winter seasons are not estimated from the microwave observations. Similar underestimation of SWE is observed in the flatland region (Figure 20), where there is no snow estimated for shorter snow events in the 2010 and 2011 winter seasons and only a small overestimation of SWE observed for 3 days in February Figure 19. Comparison of pixel snow water equivalent (SWE) estimate from H13 satellite product and SWE estimated from snow depth observations at Brand station (Vorarlberg region). Station is located approximately at the mean pixel elevation of H13 product (1014 m a.sl.).

32 Figure 20. Comparison of pixel snow water equivalent (SWE) estimate from H13 satellite product and SWE estimated from snow depth observations at Eisenstadt station (Burgenland region). Station is located approximately at the mean pixel elevation of H13 product (184 m a.sl.).

33 3 Assimilation of snow products in a conceptual hydrological model Water stored in the snow pack represents an important component of the hydrologic balance in many regions of the world, especially in mountain regions. Monitoring and modeling of snow accumulation and melt is particularly difficult in these areas because of limited availability and large spatial variability of hydrologic and climate data. Advances and increasing availability of remote sensing observations is appealing for hydrologic applications as they provide an alternative data source with adequate spatial and temporal resolution. This part of the report describes the methodology and evaluation of assimilation of MSG- SEVIRI snow cover data into calibration of a conceptual hydrologic model. The methodology follows the approach of Parajka and Blöschl (2008), but applies MSG-SEVIRI (H10) dataset for validating and multiple objective calibration of a hydrologic model. The calibration and validation runoff and snow model performances are evaluated and compared in the calibration and validation periods. Such assessment is an important step in developing a strategy for diagnostic assimilation of snow cover data in operational runoff predictions. 3.1 Assimilation of satellite data in model calibration Hydrologic model The hydrologic model tested for the MSG-SEVIRI (H10) snow product integration is a semidistributed conceptual rainfall runoff model, following the structure of the HBV model (Bergström, 1976) and uses elevations zones of 200 m. The model runs on a daily time step and consists of snow, soil moisture and flow routing routines. The snow routine simulates snow accumulation and melt using a concept of threshold air temperatures and a simple degree-day melting approach. Mean daily precipitation in an elevation zone is partitioned into rain and snow, based on the mean daily air temperature and the rain (TR) and snow (TS) air temperature thresholds. The catch deficit of the precipitation gauges during snowfall is corrected by a snow correction factor (CSF). Snow accumulation starts at air temperatures below a melt air temperature threshold (TM). The amount of water stored in a snow pack is described by the snow water equivalent (SWE), which is a state variable of the model and is simulated independently in each elevation zone of a catchment. Snow melt starts at air temperatures above a TM threshold and is proportional to a degree day factor (DDF) and the

34 difference between air temperature and a TM threshold. The soil moisture routine represents runoff generation and changes in the soil moisture state of the catchment. It is characterised by three model parameters: maximum soil moisture storage (FC), soil moisture state above which evaporation is at its potential rate (LP) and a parameter relating runoff generation to the soil moisture state (BETA). Runoff routing on the hillslopes is represented by an upper and a lower soil reservoir. Excess rainfall enters the upper zone reservoir and leaves this reservoir through three paths, outflow from the reservoir based on a fast storage coefficient (K1); percolation to the lower zone with a constant percolation rate (CP); and, if a threshold of the storage state (LSUZ) is exceeded, through an additional outlet based on a very fast storage coefficient (K0). Water leaves the lower zone based on a slow storage coefficient (K2). The outflow from both reservoirs is then routed by a triangular transfer function using a free model parameter (CR). Table 6. Hydrologic model parameters and lower (p l ) and upper (p u ) bounds used in model calibration. Model parameter j Model component p l p u CSF [-] Snow TM Snow DDF [mm/ C day] Snow LP/FC [-] Soil FC [mm] Soil B [-] Soil K 0 [days] Runoff K 1 [days] Runoff K 2 [days] Runoff C P [mm/day] Runoff LS UZ [mm] Runoff

35 From a total of 14 model parameters, 3 parameters were fixed (TR=2ºC, TS=-2ºC, CR=26.5, for details see e.g. p.5 and Figure 6 of Parajka et al. 2007b) and 11 parameters (Table 6) were estimated by automatic model calibration. More detailed information about the model structure and the model equations are given in the appendix of Parajka et al. (2007a); and examples of its application in hydrological modelling in Austria is presented, e.g., in Parajka et al. (2005a, 2005b, 2007b, 2008) Efficiency and error measures for runoff and snow covered area Calibration and validation of the model is based on a number of efficiency measures and error measures that represent the match (or mismatch) of the simulation and the data. For runoff, the Nash-Sutcliffe Model efficiency has been used in two variants, emphasize high and low flows, respectively: M and M E log E, that M E i= 1 = n n ( Qobs, i Qsim, i ) ( Q i Qobs ) obs, i= (1) and M log E i= 1 = n n ( log( Qobs, i ) log( Qsim, i )) ( log( Q i Qobs ) obs, ) log( ) i= (2) where Q sim, i is the simulated runoff on day i, obs i Q, is the observed runoff, Q obs is the average of the observed runoff over the calibration (or verification) period of n days. Also a relative volume error V E of runoff has been analysed: V E n Q i= 1 sim, i i= 1 i= 1 = n Q n obs, i Q obs, i (3) For snow covered area, the comparison is less straightforward as the model is based on elevation zones while satellite snow data (MSG-SEVIRI) are raster data. A schematic example of SWE simulation for a hypothetical catchment with four elevation zones (A, B, C, D) is

36 presented in Figure 20. This example shows that the model simulates a uniform distribution of SWE in each elevation zone, which is in contrast with the gridded representation of satellite snow cover map (right panel of Figure 20). Another difference between these two snow representations stems from the fact that the model simulates the amount (volume) of water stored in the form of snow, while the satellite image shows only whether the spatial unit of the snow mapping (pixel) is covered by snow, land or is classified as missing information (mostly representing the clouds). This indicates that comparison of satellite snow cover data with the SWE model simulations is possible only in an indirect way. Figure 21. Schematic comparison of simulated snow water equivalent SWE (left) and satellite snow cover (right). The A, B, C and D polygons represent the elevation zones of a catchment. Both maps are illustrative examples. The comparison is performed in individual elevation zones of a catchment. Two types of snow errors are evaluated. The first, termed model overestimation error ( S ), counts the number of days m O when the hydrologic model simulates zone SWE greater than a threshold but satellite image indicates that no snow is present in the zone, i.e.: O E S O E = 1 m l l j= 1 m ( SWE > ξ O SWE ) ( SCA = 0) (4)

37 where SWE is the simulated snow water equivalent in one zone, SCA is the snow covered area derived from a satellite within this zone, m is the number of days where satellite images are available (with cloud cover less than a threshold ξ C ), l is the number of zones of a particular catchment, and ξ SWE is a threshold that determines when a zone can be essentially considered snow free in terms of the simulations. An example of a day that would contribute to the snow overestimation error is presented in zone C of Figure 20. The second error, termed model underestimation error ( S ) counts the number of days when the hydrologic model does not simulate snow in a zone but satellite indicates that snow covered area greater than a threshold is present in the zone, i.e.: U E m U S U E 1 = m l l j= 1 m ( SWE = 0) ( SCA > ξ ) (5) U SCA where ξ SCA is a threshold that determines when a zone can be essentially considered snow free in terms of the satellite data. An example of a day that would contribute to the snow underestimation error is presented in zone A of Figure 20. The percent or fraction of snow covered area, SCA, within each zone was calculated from the satellite data as: S SCA = (6) S + L where S and L represent the number of pixels mapped as snow and land, respectively, for a given day and a given zone. The reliability and accuracy of the SCA estimation depends on the spatial extent of clouds occurring in an elevation zone. Only those days of the SCA images were hence used for a particular day and elevation zone if the cloud coverage was less than a threshold ξ C : C < ξ C (7) where C is the fractional cloud cover for a particular day and elevation zone. The thresholds ξ SWE, ξ SCA and ξ C have been chosen on the basis of a sensitivity analysis performed in Parajka et al. (2008). Based on this assessment, thresholds of ξ C =60%, ξ SCA =25% and ξ SWE =0 were selected for the model snow error evaluations.

38 As no satellite data are available in the verification period, ground based snow depth observations were applied for the validation of the snow model instead. The model errors O S D and U S D are defined in a similar way as in Eq. 4 and 5 but instead of using satellite SCA, the ground snow depth data were spatially interpolated from which the snow covered area was calculated for each elevation zone. The pixels were mapped as snow covered when the interpolated snow depth exceeded 1 cm and considered as land otherwise. Snow depth maps were interpolated by the external drift kriging method, using elevation as auxiliary variable Calibration to runoff alone In a first variant, termed single-objective calibration, we emulate the usual model calibration and estimate the parameters of the hydrologic model using measured runoff only. The runoff objective function is defined as: Z Q Q log ( M ) + (1 w ) (1 M ) = w (8) 1 E Q E where the weight w Q is set to 0.5. The idea of Eq. 8 is to combine two agreement measures M and M E log E, that emphasize high and low flows, respectively. The SCE-UA automatic calibration procedure (Duan et al, 1992) is used to minimize Eq. 8. No snow satellite data are used for the calibration in this variant but they are used for assessing the errors of the snow simulations by analyzing the O S E and U S E errors for all catchments Calibration to both runoff and satellite snow cover In a second variant, termed multiple-objective calibration, we use both runoff data and satellite snow cover data (MSG-SEVIRI H10) to calibrate the model by minimizing a compound objective function the runoff and the snow cover, respectively: Z M, which involves two parts Z Q and Z S that are related to Z M = w Z + ( 1 w ) Z (9) S S S S

39 The w S is chosen on the basis of sensitivity analyses (see section 4.1). The snow part Z S of the compound objective function represents the sum of the over- and underestimation snow errors: Z S O U = w1 SE + w2 SE (10) which were equally weighted in this study, so w 1 and w 2 were both set to 1.0. The same calibration and verification periods were used in the two variants of model calibration. 3.2 Data The integration of satellite snow data (MSG-SEVIRI) into a conceptual hydrologic model is tested and evaluated in 148 catchments in Austria (Figure 1, Table 1). These catchments are located in different physiographic and climatic zones and have different sizes, ranging from 25 km² to 9770 km² with a median size of 369 km². Elevations of the study region range from 115 m a.s.l. to 3797 m a.s.l.. Mean annual precipitation ranges from less than 400 mm/year in the East to almost 3000 mm/year in the West. Land use is mainly agricultural in the lowlands and forest in the medium elevation ranges. Alpine vegetation and rocks prevail in the highest mountain regions. Such diverse physiographic and landscape characteristics suggest that the study region is representative of a wider spatial domain and the results may be applicable in catchments with similar characteristics. The hydrologic data set used in this study includes runoff data of the 144 catchments to calibrate and validate the hydrological model for different periods (Figure 21). The data also include daily precipitation at 1091 stations and daily air temperature at 240 climatic stations as an input to the hydrological model. The precipitation data were spatially interpolated by external drift kriging and the air temperature data were interpolated by the least-squares trend prediction method (Pebesma, 2001), using elevation as an auxiliary variable in both cases. To validate the snow model when no satellite snow cover data were available, ground snow depth data at 1091 stations were also used.

40 Figure 22. Location of 144 runoff gauges (circles) used in the analysis. The dataset used in this study consists of two parts. The first is a calibration dataset, which includes the hydrologic and satellite data in the period from January, 2007 to December, The second is a verification dataset, which includes the hydrologic data in the period from November 1, 1976 to December 31, In the verification period, satellite data are not available for the entire period, thus ground based snow depth observations were applied in validation instead. 3.3 Results Model performance - calibration to runoff alone The efficiency of the hydrologic model to simulate runoff and snow is evaluated in Table 7. The assessment of model performance represents a typical modelling concept where only runoff data are available for hydrologic model calibration. The runoff and snow model efficiencies are summarized over 144 catchments separately for the calibration and verification periods. The medians of the calibration runoff efficiencies ME and MElog are 0.81 and 0.87, respectively, which indicates a good overall agreement between observed and simulated runoff. (See Merz and Blöschl, 2004 for an assessment of what is considered a good model performance.) The median runoff volume error (VE) is -1.4%, which indicates that the calibration is essentially unbiased. The comparison between

41 simulated snow and observed snow cover from MSG-SEVIRI indicates that the median model O E overestimation ( S ) and underestimation ( S U E ) errors is 15% and 20%, respectively. This means that snow model errors in an individual elevation zone tend to last 15-20% of cloud-free days of the calibration period. The cloud-free zone/days represent, in this case, approximately 50% of the calibration period. The snow model performance evaluated against interpolated snow depth ( S, S ) shows that the hydrologic model tends to underestimate the snow cover derived from O D U D interpolated snow depth observations. The median of snow depth overestimation and underestimation errors is 1.4% and 7.8%, respectively. The spatial pattern of runoff and snow model performance is presented in Figure 23. The top map shows that the model performs very well in the mountain regions, where the ME runoff exceeds 0.80 in most catchments. Somewhat lower ME ( ) is obtained in the flatland regions of Northern and South-Eastern part of Austria. The snow model performance patterns (bottom map) indicates that the error in the mountains tend to be larger than in the flatlands. The largest snow errors with respect to MSG-SEVIRI exceed 40% in the Central part of Austria.

42 Figure 23. Runoff (ME) and MSG-SEVIRI snow (ZS) model efficiency obtained by single-objective calibration to measured runoff only in the calibration period